Thu, 31 Mar 2011 13:22:34 -0700
7032696: Fix for 7029152 broke VM
Summary: StrIntrinsicNode::Ideal() should not optimize memory during Parse.
Reviewed-by: jrose, never
1 /*
2 * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/connode.hpp"
31 #include "opto/divnode.hpp"
32 #include "opto/loopnode.hpp"
33 #include "opto/mulnode.hpp"
34 #include "opto/rootnode.hpp"
35 #include "opto/runtime.hpp"
36 #include "opto/subnode.hpp"
38 //------------------------------is_loop_exit-----------------------------------
39 // Given an IfNode, return the loop-exiting projection or NULL if both
40 // arms remain in the loop.
41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
43 PhaseIdealLoop *phase = _phase;
44 // Test is an IfNode, has 2 projections. If BOTH are in the loop
45 // we need loop unswitching instead of peeling.
46 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
47 return iff->raw_out(0);
48 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
49 return iff->raw_out(1);
50 return NULL;
51 }
54 //=============================================================================
57 //------------------------------record_for_igvn----------------------------
58 // Put loop body on igvn work list
59 void IdealLoopTree::record_for_igvn() {
60 for( uint i = 0; i < _body.size(); i++ ) {
61 Node *n = _body.at(i);
62 _phase->_igvn._worklist.push(n);
63 }
64 }
66 //------------------------------compute_profile_trip_cnt----------------------------
67 // Compute loop trip count from profile data as
68 // (backedge_count + loop_exit_count) / loop_exit_count
69 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
70 if (!_head->is_CountedLoop()) {
71 return;
72 }
73 CountedLoopNode* head = _head->as_CountedLoop();
74 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
75 return; // Already computed
76 }
77 float trip_cnt = (float)max_jint; // default is big
79 Node* back = head->in(LoopNode::LoopBackControl);
80 while (back != head) {
81 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
82 back->in(0) &&
83 back->in(0)->is_If() &&
84 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
85 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
86 break;
87 }
88 back = phase->idom(back);
89 }
90 if (back != head) {
91 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
92 back->in(0), "if-projection exists");
93 IfNode* back_if = back->in(0)->as_If();
94 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
96 // Now compute a loop exit count
97 float loop_exit_cnt = 0.0f;
98 for( uint i = 0; i < _body.size(); i++ ) {
99 Node *n = _body[i];
100 if( n->is_If() ) {
101 IfNode *iff = n->as_If();
102 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
103 Node *exit = is_loop_exit(iff);
104 if( exit ) {
105 float exit_prob = iff->_prob;
106 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
107 if (exit_prob > PROB_MIN) {
108 float exit_cnt = iff->_fcnt * exit_prob;
109 loop_exit_cnt += exit_cnt;
110 }
111 }
112 }
113 }
114 }
115 if (loop_exit_cnt > 0.0f) {
116 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
117 } else {
118 // No exit count so use
119 trip_cnt = loop_back_cnt;
120 }
121 }
122 #ifndef PRODUCT
123 if (TraceProfileTripCount) {
124 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
125 }
126 #endif
127 head->set_profile_trip_cnt(trip_cnt);
128 }
130 //---------------------is_invariant_addition-----------------------------
131 // Return nonzero index of invariant operand for an Add or Sub
132 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
133 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
134 int op = n->Opcode();
135 if (op == Op_AddI || op == Op_SubI) {
136 bool in1_invar = this->is_invariant(n->in(1));
137 bool in2_invar = this->is_invariant(n->in(2));
138 if (in1_invar && !in2_invar) return 1;
139 if (!in1_invar && in2_invar) return 2;
140 }
141 return 0;
142 }
144 //---------------------reassociate_add_sub-----------------------------
145 // Reassociate invariant add and subtract expressions:
146 //
147 // inv1 + (x + inv2) => ( inv1 + inv2) + x
148 // (x + inv2) + inv1 => ( inv1 + inv2) + x
149 // inv1 + (x - inv2) => ( inv1 - inv2) + x
150 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
151 // (x + inv2) - inv1 => (-inv1 + inv2) + x
152 // (x - inv2) + inv1 => ( inv1 - inv2) + x
153 // (x - inv2) - inv1 => (-inv1 - inv2) + x
154 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
155 // inv1 - (x - inv2) => ( inv1 + inv2) - x
156 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
157 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
158 // inv1 - (x + inv2) => ( inv1 - inv2) - x
159 //
160 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
161 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
162 if (is_invariant(n1)) return NULL;
163 int inv1_idx = is_invariant_addition(n1, phase);
164 if (!inv1_idx) return NULL;
165 // Don't mess with add of constant (igvn moves them to expression tree root.)
166 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
167 Node* inv1 = n1->in(inv1_idx);
168 Node* n2 = n1->in(3 - inv1_idx);
169 int inv2_idx = is_invariant_addition(n2, phase);
170 if (!inv2_idx) return NULL;
171 Node* x = n2->in(3 - inv2_idx);
172 Node* inv2 = n2->in(inv2_idx);
174 bool neg_x = n2->is_Sub() && inv2_idx == 1;
175 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
176 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
177 if (n1->is_Sub() && inv1_idx == 1) {
178 neg_x = !neg_x;
179 neg_inv2 = !neg_inv2;
180 }
181 Node* inv1_c = phase->get_ctrl(inv1);
182 Node* inv2_c = phase->get_ctrl(inv2);
183 Node* n_inv1;
184 if (neg_inv1) {
185 Node *zero = phase->_igvn.intcon(0);
186 phase->set_ctrl(zero, phase->C->root());
187 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
188 phase->register_new_node(n_inv1, inv1_c);
189 } else {
190 n_inv1 = inv1;
191 }
192 Node* inv;
193 if (neg_inv2) {
194 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
195 } else {
196 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
197 }
198 phase->register_new_node(inv, phase->get_early_ctrl(inv));
200 Node* addx;
201 if (neg_x) {
202 addx = new (phase->C, 3) SubINode(inv, x);
203 } else {
204 addx = new (phase->C, 3) AddINode(x, inv);
205 }
206 phase->register_new_node(addx, phase->get_ctrl(x));
207 phase->_igvn.replace_node(n1, addx);
208 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
209 _body.yank(n1);
210 return addx;
211 }
213 //---------------------reassociate_invariants-----------------------------
214 // Reassociate invariant expressions:
215 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
216 for (int i = _body.size() - 1; i >= 0; i--) {
217 Node *n = _body.at(i);
218 for (int j = 0; j < 5; j++) {
219 Node* nn = reassociate_add_sub(n, phase);
220 if (nn == NULL) break;
221 n = nn; // again
222 };
223 }
224 }
226 //------------------------------policy_peeling---------------------------------
227 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
228 // make some loop-invariant test (usually a null-check) happen before the loop.
229 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
230 Node *test = ((IdealLoopTree*)this)->tail();
231 int body_size = ((IdealLoopTree*)this)->_body.size();
232 int uniq = phase->C->unique();
233 // Peeling does loop cloning which can result in O(N^2) node construction
234 if( body_size > 255 /* Prevent overflow for large body_size */
235 || (body_size * body_size + uniq > MaxNodeLimit) ) {
236 return false; // too large to safely clone
237 }
238 while( test != _head ) { // Scan till run off top of loop
239 if( test->is_If() ) { // Test?
240 Node *ctrl = phase->get_ctrl(test->in(1));
241 if (ctrl->is_top())
242 return false; // Found dead test on live IF? No peeling!
243 // Standard IF only has one input value to check for loop invariance
244 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
245 // Condition is not a member of this loop?
246 if( !is_member(phase->get_loop(ctrl)) &&
247 is_loop_exit(test) )
248 return true; // Found reason to peel!
249 }
250 // Walk up dominators to loop _head looking for test which is
251 // executed on every path thru loop.
252 test = phase->idom(test);
253 }
254 return false;
255 }
257 //------------------------------peeled_dom_test_elim---------------------------
258 // If we got the effect of peeling, either by actually peeling or by making
259 // a pre-loop which must execute at least once, we can remove all
260 // loop-invariant dominated tests in the main body.
261 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
262 bool progress = true;
263 while( progress ) {
264 progress = false; // Reset for next iteration
265 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
266 Node *test = prev->in(0);
267 while( test != loop->_head ) { // Scan till run off top of loop
269 int p_op = prev->Opcode();
270 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
271 test->is_If() && // Test?
272 !test->in(1)->is_Con() && // And not already obvious?
273 // Condition is not a member of this loop?
274 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
275 // Walk loop body looking for instances of this test
276 for( uint i = 0; i < loop->_body.size(); i++ ) {
277 Node *n = loop->_body.at(i);
278 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
279 // IfNode was dominated by version in peeled loop body
280 progress = true;
281 dominated_by( old_new[prev->_idx], n );
282 }
283 }
284 }
285 prev = test;
286 test = idom(test);
287 } // End of scan tests in loop
289 } // End of while( progress )
290 }
292 //------------------------------do_peeling-------------------------------------
293 // Peel the first iteration of the given loop.
294 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
295 // The pre-loop illegally has 2 control users (old & new loops).
296 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
297 // Do this by making the old-loop fall-in edges act as if they came
298 // around the loopback from the prior iteration (follow the old-loop
299 // backedges) and then map to the new peeled iteration. This leaves
300 // the pre-loop with only 1 user (the new peeled iteration), but the
301 // peeled-loop backedge has 2 users.
302 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
303 // extra backedge user.
304 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
306 C->set_major_progress();
307 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
308 // 'pre' loop from the main and the 'pre' can no longer have it's
309 // iterations adjusted. Therefore, we need to declare this loop as
310 // no longer a 'main' loop; it will need new pre and post loops before
311 // we can do further RCE.
312 #ifndef PRODUCT
313 if (TraceLoopOpts) {
314 tty->print("Peel ");
315 loop->dump_head();
316 }
317 #endif
318 Node *h = loop->_head;
319 if (h->is_CountedLoop()) {
320 CountedLoopNode *cl = h->as_CountedLoop();
321 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
322 cl->set_trip_count(cl->trip_count() - 1);
323 if (cl->is_main_loop()) {
324 cl->set_normal_loop();
325 #ifndef PRODUCT
326 if (PrintOpto && VerifyLoopOptimizations) {
327 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
328 loop->dump_head();
329 }
330 #endif
331 }
332 }
334 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
335 // The pre-loop illegally has 2 control users (old & new loops).
336 clone_loop( loop, old_new, dom_depth(loop->_head) );
339 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
340 // Do this by making the old-loop fall-in edges act as if they came
341 // around the loopback from the prior iteration (follow the old-loop
342 // backedges) and then map to the new peeled iteration. This leaves
343 // the pre-loop with only 1 user (the new peeled iteration), but the
344 // peeled-loop backedge has 2 users.
345 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) {
346 Node* old = loop->_head->fast_out(j);
347 if( old->in(0) == loop->_head && old->req() == 3 &&
348 (old->is_Loop() || old->is_Phi()) ) {
349 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
350 if( !new_exit_value ) // Backedge value is ALSO loop invariant?
351 // Then loop body backedge value remains the same.
352 new_exit_value = old->in(LoopNode::LoopBackControl);
353 _igvn.hash_delete(old);
354 old->set_req(LoopNode::EntryControl, new_exit_value);
355 }
356 }
359 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
360 // extra backedge user.
361 Node *nnn = old_new[loop->_head->_idx];
362 _igvn.hash_delete(nnn);
363 nnn->set_req(LoopNode::LoopBackControl, C->top());
364 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) {
365 Node* use = nnn->fast_out(j2);
366 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) {
367 _igvn.hash_delete(use);
368 use->set_req(LoopNode::LoopBackControl, C->top());
369 }
370 }
373 // Step 4: Correct dom-depth info. Set to loop-head depth.
374 int dd = dom_depth(loop->_head);
375 set_idom(loop->_head, loop->_head->in(1), dd);
376 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
377 Node *old = loop->_body.at(j3);
378 Node *nnn = old_new[old->_idx];
379 if (!has_ctrl(nnn))
380 set_idom(nnn, idom(nnn), dd-1);
381 // While we're at it, remove any SafePoints from the peeled code
382 if( old->Opcode() == Op_SafePoint ) {
383 Node *nnn = old_new[old->_idx];
384 lazy_replace(nnn,nnn->in(TypeFunc::Control));
385 }
386 }
388 // Now force out all loop-invariant dominating tests. The optimizer
389 // finds some, but we _know_ they are all useless.
390 peeled_dom_test_elim(loop,old_new);
392 loop->record_for_igvn();
393 }
395 //------------------------------policy_maximally_unroll------------------------
396 // Return exact loop trip count, or 0 if not maximally unrolling
397 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
398 CountedLoopNode *cl = _head->as_CountedLoop();
399 assert(cl->is_normal_loop(), "");
401 Node *init_n = cl->init_trip();
402 Node *limit_n = cl->limit();
404 // Non-constant bounds
405 if (init_n == NULL || !init_n->is_Con() ||
406 limit_n == NULL || !limit_n->is_Con() ||
407 // protect against stride not being a constant
408 !cl->stride_is_con()) {
409 return false;
410 }
411 int init = init_n->get_int();
412 int limit = limit_n->get_int();
413 int span = limit - init;
414 int stride = cl->stride_con();
416 if (init >= limit || stride > span) {
417 // return a false (no maximally unroll) and the regular unroll/peel
418 // route will make a small mess which CCP will fold away.
419 return false;
420 }
421 uint trip_count = span/stride; // trip_count can be greater than 2 Gig.
422 assert( (int)trip_count*stride == span, "must divide evenly" );
424 // Real policy: if we maximally unroll, does it get too big?
425 // Allow the unrolled mess to get larger than standard loop
426 // size. After all, it will no longer be a loop.
427 uint body_size = _body.size();
428 uint unroll_limit = (uint)LoopUnrollLimit * 4;
429 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
430 cl->set_trip_count(trip_count);
431 if (trip_count > unroll_limit || body_size > unroll_limit) {
432 return false;
433 }
435 // Currently we don't have policy to optimize one iteration loops.
436 // Maximally unrolling transformation is used for that:
437 // it is peeled and the original loop become non reachable (dead).
438 if (trip_count == 1)
439 return true;
441 // Do not unroll a loop with String intrinsics code.
442 // String intrinsics are large and have loops.
443 for (uint k = 0; k < _body.size(); k++) {
444 Node* n = _body.at(k);
445 switch (n->Opcode()) {
446 case Op_StrComp:
447 case Op_StrEquals:
448 case Op_StrIndexOf:
449 case Op_AryEq: {
450 return false;
451 }
452 } // switch
453 }
455 if (body_size <= unroll_limit) {
456 uint new_body_size = body_size * trip_count;
457 if (new_body_size <= unroll_limit &&
458 body_size == new_body_size / trip_count &&
459 // Unrolling can result in a large amount of node construction
460 new_body_size < MaxNodeLimit - phase->C->unique()) {
461 return true; // maximally unroll
462 }
463 }
465 return false; // Do not maximally unroll
466 }
469 //------------------------------policy_unroll----------------------------------
470 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
471 // the loop is a CountedLoop and the body is small enough.
472 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
474 CountedLoopNode *cl = _head->as_CountedLoop();
475 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
477 // protect against stride not being a constant
478 if (!cl->stride_is_con()) return false;
480 // protect against over-unrolling
481 if (cl->trip_count() <= 1) return false;
483 int future_unroll_ct = cl->unrolled_count() * 2;
485 // Don't unroll if the next round of unrolling would push us
486 // over the expected trip count of the loop. One is subtracted
487 // from the expected trip count because the pre-loop normally
488 // executes 1 iteration.
489 if (UnrollLimitForProfileCheck > 0 &&
490 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
491 future_unroll_ct > UnrollLimitForProfileCheck &&
492 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
493 return false;
494 }
496 // When unroll count is greater than LoopUnrollMin, don't unroll if:
497 // the residual iterations are more than 10% of the trip count
498 // and rounds of "unroll,optimize" are not making significant progress
499 // Progress defined as current size less than 20% larger than previous size.
500 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
501 future_unroll_ct > LoopUnrollMin &&
502 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
503 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
504 return false;
505 }
507 Node *init_n = cl->init_trip();
508 Node *limit_n = cl->limit();
509 // Non-constant bounds.
510 // Protect against over-unrolling when init or/and limit are not constant
511 // (so that trip_count's init value is maxint) but iv range is known.
512 if (init_n == NULL || !init_n->is_Con() ||
513 limit_n == NULL || !limit_n->is_Con()) {
514 Node* phi = cl->phi();
515 if (phi != NULL) {
516 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
517 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
518 int next_stride = cl->stride_con() * 2; // stride after this unroll
519 if (next_stride > 0) {
520 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
521 iv_type->_lo + next_stride > iv_type->_hi) {
522 return false; // over-unrolling
523 }
524 } else if (next_stride < 0) {
525 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
526 iv_type->_hi + next_stride < iv_type->_lo) {
527 return false; // over-unrolling
528 }
529 }
530 }
531 }
533 // Adjust body_size to determine if we unroll or not
534 uint body_size = _body.size();
535 // Key test to unroll CaffeineMark's Logic test
536 int xors_in_loop = 0;
537 // Also count ModL, DivL and MulL which expand mightly
538 for (uint k = 0; k < _body.size(); k++) {
539 Node* n = _body.at(k);
540 switch (n->Opcode()) {
541 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
542 case Op_ModL: body_size += 30; break;
543 case Op_DivL: body_size += 30; break;
544 case Op_MulL: body_size += 10; break;
545 case Op_StrComp:
546 case Op_StrEquals:
547 case Op_StrIndexOf:
548 case Op_AryEq: {
549 // Do not unroll a loop with String intrinsics code.
550 // String intrinsics are large and have loops.
551 return false;
552 }
553 } // switch
554 }
556 // Check for being too big
557 if (body_size > (uint)LoopUnrollLimit) {
558 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
559 // Normal case: loop too big
560 return false;
561 }
563 // Check for stride being a small enough constant
564 if (abs(cl->stride_con()) > (1<<3)) return false;
566 // Unroll once! (Each trip will soon do double iterations)
567 return true;
568 }
570 //------------------------------policy_align-----------------------------------
571 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
572 // expression that does the alignment. Note that only one array base can be
573 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
574 // if we vectorize short memory ops into longer memory ops, we may want to
575 // increase alignment.
576 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
577 return false;
578 }
580 //------------------------------policy_range_check-----------------------------
581 // Return TRUE or FALSE if the loop should be range-check-eliminated.
582 // Actually we do iteration-splitting, a more powerful form of RCE.
583 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
584 if( !RangeCheckElimination ) return false;
586 CountedLoopNode *cl = _head->as_CountedLoop();
587 // If we unrolled with no intention of doing RCE and we later
588 // changed our minds, we got no pre-loop. Either we need to
589 // make a new pre-loop, or we gotta disallow RCE.
590 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
591 Node *trip_counter = cl->phi();
593 // Check loop body for tests of trip-counter plus loop-invariant vs
594 // loop-invariant.
595 for( uint i = 0; i < _body.size(); i++ ) {
596 Node *iff = _body[i];
597 if( iff->Opcode() == Op_If ) { // Test?
599 // Comparing trip+off vs limit
600 Node *bol = iff->in(1);
601 if( bol->req() != 2 ) continue; // dead constant test
602 if (!bol->is_Bool()) {
603 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
604 continue;
605 }
606 Node *cmp = bol->in(1);
608 Node *rc_exp = cmp->in(1);
609 Node *limit = cmp->in(2);
611 Node *limit_c = phase->get_ctrl(limit);
612 if( limit_c == phase->C->top() )
613 return false; // Found dead test on live IF? No RCE!
614 if( is_member(phase->get_loop(limit_c) ) ) {
615 // Compare might have operands swapped; commute them
616 rc_exp = cmp->in(2);
617 limit = cmp->in(1);
618 limit_c = phase->get_ctrl(limit);
619 if( is_member(phase->get_loop(limit_c) ) )
620 continue; // Both inputs are loop varying; cannot RCE
621 }
623 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
624 continue;
625 }
626 // Yeah! Found a test like 'trip+off vs limit'
627 // Test is an IfNode, has 2 projections. If BOTH are in the loop
628 // we need loop unswitching instead of iteration splitting.
629 if( is_loop_exit(iff) )
630 return true; // Found reason to split iterations
631 } // End of is IF
632 }
634 return false;
635 }
637 //------------------------------policy_peel_only-------------------------------
638 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
639 // for unrolling loops with NO array accesses.
640 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
642 for( uint i = 0; i < _body.size(); i++ )
643 if( _body[i]->is_Mem() )
644 return false;
646 // No memory accesses at all!
647 return true;
648 }
650 //------------------------------clone_up_backedge_goo--------------------------
651 // If Node n lives in the back_ctrl block and cannot float, we clone a private
652 // version of n in preheader_ctrl block and return that, otherwise return n.
653 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
654 if( get_ctrl(n) != back_ctrl ) return n;
656 Node *x = NULL; // If required, a clone of 'n'
657 // Check for 'n' being pinned in the backedge.
658 if( n->in(0) && n->in(0) == back_ctrl ) {
659 x = n->clone(); // Clone a copy of 'n' to preheader
660 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
661 }
663 // Recursive fixup any other input edges into x.
664 // If there are no changes we can just return 'n', otherwise
665 // we need to clone a private copy and change it.
666 for( uint i = 1; i < n->req(); i++ ) {
667 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
668 if( g != n->in(i) ) {
669 if( !x )
670 x = n->clone();
671 x->set_req(i, g);
672 }
673 }
674 if( x ) { // x can legally float to pre-header location
675 register_new_node( x, preheader_ctrl );
676 return x;
677 } else { // raise n to cover LCA of uses
678 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
679 }
680 return n;
681 }
683 //------------------------------insert_pre_post_loops--------------------------
684 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
685 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
686 // alignment. Useful to unroll loops that do no array accesses.
687 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
689 #ifndef PRODUCT
690 if (TraceLoopOpts) {
691 if (peel_only)
692 tty->print("PeelMainPost ");
693 else
694 tty->print("PreMainPost ");
695 loop->dump_head();
696 }
697 #endif
698 C->set_major_progress();
700 // Find common pieces of the loop being guarded with pre & post loops
701 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
702 assert( main_head->is_normal_loop(), "" );
703 CountedLoopEndNode *main_end = main_head->loopexit();
704 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
705 uint dd_main_head = dom_depth(main_head);
706 uint max = main_head->outcnt();
708 Node *pre_header= main_head->in(LoopNode::EntryControl);
709 Node *init = main_head->init_trip();
710 Node *incr = main_end ->incr();
711 Node *limit = main_end ->limit();
712 Node *stride = main_end ->stride();
713 Node *cmp = main_end ->cmp_node();
714 BoolTest::mask b_test = main_end->test_trip();
716 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
717 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
718 if( bol->outcnt() != 1 ) {
719 bol = bol->clone();
720 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
721 _igvn.hash_delete(main_end);
722 main_end->set_req(CountedLoopEndNode::TestValue, bol);
723 }
724 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
725 if( cmp->outcnt() != 1 ) {
726 cmp = cmp->clone();
727 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
728 _igvn.hash_delete(bol);
729 bol->set_req(1, cmp);
730 }
732 //------------------------------
733 // Step A: Create Post-Loop.
734 Node* main_exit = main_end->proj_out(false);
735 assert( main_exit->Opcode() == Op_IfFalse, "" );
736 int dd_main_exit = dom_depth(main_exit);
738 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
739 // loop pre-header illegally has 2 control users (old & new loops).
740 clone_loop( loop, old_new, dd_main_exit );
741 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
742 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
743 post_head->set_post_loop(main_head);
745 // Reduce the post-loop trip count.
746 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
747 post_end->_prob = PROB_FAIR;
749 // Build the main-loop normal exit.
750 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
751 _igvn.register_new_node_with_optimizer( new_main_exit );
752 set_idom(new_main_exit, main_end, dd_main_exit );
753 set_loop(new_main_exit, loop->_parent);
755 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
756 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
757 // (the main-loop trip-counter exit value) because we will be changing
758 // the exit value (via unrolling) so we cannot constant-fold away the zero
759 // trip guard until all unrolling is done.
760 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
761 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
762 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
763 register_new_node( zer_opaq, new_main_exit );
764 register_new_node( zer_cmp , new_main_exit );
765 register_new_node( zer_bol , new_main_exit );
767 // Build the IfNode
768 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
769 _igvn.register_new_node_with_optimizer( zer_iff );
770 set_idom(zer_iff, new_main_exit, dd_main_exit);
771 set_loop(zer_iff, loop->_parent);
773 // Plug in the false-path, taken if we need to skip post-loop
774 _igvn.hash_delete( main_exit );
775 main_exit->set_req(0, zer_iff);
776 _igvn._worklist.push(main_exit);
777 set_idom(main_exit, zer_iff, dd_main_exit);
778 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
779 // Make the true-path, must enter the post loop
780 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
781 _igvn.register_new_node_with_optimizer( zer_taken );
782 set_idom(zer_taken, zer_iff, dd_main_exit);
783 set_loop(zer_taken, loop->_parent);
784 // Plug in the true path
785 _igvn.hash_delete( post_head );
786 post_head->set_req(LoopNode::EntryControl, zer_taken);
787 set_idom(post_head, zer_taken, dd_main_exit);
789 // Step A3: Make the fall-in values to the post-loop come from the
790 // fall-out values of the main-loop.
791 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
792 Node* main_phi = main_head->fast_out(i);
793 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
794 Node *post_phi = old_new[main_phi->_idx];
795 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
796 post_head->init_control(),
797 main_phi->in(LoopNode::LoopBackControl));
798 _igvn.hash_delete(post_phi);
799 post_phi->set_req( LoopNode::EntryControl, fallmain );
800 }
801 }
803 // Update local caches for next stanza
804 main_exit = new_main_exit;
807 //------------------------------
808 // Step B: Create Pre-Loop.
810 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
811 // loop pre-header illegally has 2 control users (old & new loops).
812 clone_loop( loop, old_new, dd_main_head );
813 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
814 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
815 pre_head->set_pre_loop(main_head);
816 Node *pre_incr = old_new[incr->_idx];
818 // Reduce the pre-loop trip count.
819 pre_end->_prob = PROB_FAIR;
821 // Find the pre-loop normal exit.
822 Node* pre_exit = pre_end->proj_out(false);
823 assert( pre_exit->Opcode() == Op_IfFalse, "" );
824 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
825 _igvn.register_new_node_with_optimizer( new_pre_exit );
826 set_idom(new_pre_exit, pre_end, dd_main_head);
827 set_loop(new_pre_exit, loop->_parent);
829 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
830 // pre-loop, the main-loop may not execute at all. Later in life this
831 // zero-trip guard will become the minimum-trip guard when we unroll
832 // the main-loop.
833 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
834 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
835 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
836 register_new_node( min_opaq, new_pre_exit );
837 register_new_node( min_cmp , new_pre_exit );
838 register_new_node( min_bol , new_pre_exit );
840 // Build the IfNode (assume the main-loop is executed always).
841 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
842 _igvn.register_new_node_with_optimizer( min_iff );
843 set_idom(min_iff, new_pre_exit, dd_main_head);
844 set_loop(min_iff, loop->_parent);
846 // Plug in the false-path, taken if we need to skip main-loop
847 _igvn.hash_delete( pre_exit );
848 pre_exit->set_req(0, min_iff);
849 set_idom(pre_exit, min_iff, dd_main_head);
850 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
851 // Make the true-path, must enter the main loop
852 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
853 _igvn.register_new_node_with_optimizer( min_taken );
854 set_idom(min_taken, min_iff, dd_main_head);
855 set_loop(min_taken, loop->_parent);
856 // Plug in the true path
857 _igvn.hash_delete( main_head );
858 main_head->set_req(LoopNode::EntryControl, min_taken);
859 set_idom(main_head, min_taken, dd_main_head);
861 // Step B3: Make the fall-in values to the main-loop come from the
862 // fall-out values of the pre-loop.
863 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
864 Node* main_phi = main_head->fast_out(i2);
865 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
866 Node *pre_phi = old_new[main_phi->_idx];
867 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
868 main_head->init_control(),
869 pre_phi->in(LoopNode::LoopBackControl));
870 _igvn.hash_delete(main_phi);
871 main_phi->set_req( LoopNode::EntryControl, fallpre );
872 }
873 }
875 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
876 // RCE and alignment may change this later.
877 Node *cmp_end = pre_end->cmp_node();
878 assert( cmp_end->in(2) == limit, "" );
879 Node *pre_limit = new (C, 3) AddINode( init, stride );
881 // Save the original loop limit in this Opaque1 node for
882 // use by range check elimination.
883 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
885 register_new_node( pre_limit, pre_head->in(0) );
886 register_new_node( pre_opaq , pre_head->in(0) );
888 // Since no other users of pre-loop compare, I can hack limit directly
889 assert( cmp_end->outcnt() == 1, "no other users" );
890 _igvn.hash_delete(cmp_end);
891 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
893 // Special case for not-equal loop bounds:
894 // Change pre loop test, main loop test, and the
895 // main loop guard test to use lt or gt depending on stride
896 // direction:
897 // positive stride use <
898 // negative stride use >
900 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
902 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
903 // Modify pre loop end condition
904 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
905 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
906 register_new_node( new_bol0, pre_head->in(0) );
907 _igvn.hash_delete(pre_end);
908 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
909 // Modify main loop guard condition
910 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
911 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
912 register_new_node( new_bol1, new_pre_exit );
913 _igvn.hash_delete(min_iff);
914 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
915 // Modify main loop end condition
916 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
917 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
918 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
919 _igvn.hash_delete(main_end);
920 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
921 }
923 // Flag main loop
924 main_head->set_main_loop();
925 if( peel_only ) main_head->set_main_no_pre_loop();
927 // It's difficult to be precise about the trip-counts
928 // for the pre/post loops. They are usually very short,
929 // so guess that 4 trips is a reasonable value.
930 post_head->set_profile_trip_cnt(4.0);
931 pre_head->set_profile_trip_cnt(4.0);
933 // Now force out all loop-invariant dominating tests. The optimizer
934 // finds some, but we _know_ they are all useless.
935 peeled_dom_test_elim(loop,old_new);
936 }
938 //------------------------------is_invariant-----------------------------
939 // Return true if n is invariant
940 bool IdealLoopTree::is_invariant(Node* n) const {
941 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
942 if (n_c->is_top()) return false;
943 return !is_member(_phase->get_loop(n_c));
944 }
947 //------------------------------do_unroll--------------------------------------
948 // Unroll the loop body one step - make each trip do 2 iterations.
949 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
950 assert(LoopUnrollLimit, "");
951 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
952 CountedLoopEndNode *loop_end = loop_head->loopexit();
953 assert(loop_end, "");
954 #ifndef PRODUCT
955 if (PrintOpto && VerifyLoopOptimizations) {
956 tty->print("Unrolling ");
957 loop->dump_head();
958 } else if (TraceLoopOpts) {
959 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
960 loop->dump_head();
961 }
962 #endif
964 // Remember loop node count before unrolling to detect
965 // if rounds of unroll,optimize are making progress
966 loop_head->set_node_count_before_unroll(loop->_body.size());
968 Node *ctrl = loop_head->in(LoopNode::EntryControl);
969 Node *limit = loop_head->limit();
970 Node *init = loop_head->init_trip();
971 Node *stride = loop_head->stride();
973 Node *opaq = NULL;
974 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
975 assert( loop_head->is_main_loop(), "" );
976 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
977 Node *iff = ctrl->in(0);
978 assert( iff->Opcode() == Op_If, "" );
979 Node *bol = iff->in(1);
980 assert( bol->Opcode() == Op_Bool, "" );
981 Node *cmp = bol->in(1);
982 assert( cmp->Opcode() == Op_CmpI, "" );
983 opaq = cmp->in(2);
984 // Occasionally it's possible for a pre-loop Opaque1 node to be
985 // optimized away and then another round of loop opts attempted.
986 // We can not optimize this particular loop in that case.
987 if( opaq->Opcode() != Op_Opaque1 )
988 return; // Cannot find pre-loop! Bail out!
989 }
991 C->set_major_progress();
993 // Adjust max trip count. The trip count is intentionally rounded
994 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
995 // the main, unrolled, part of the loop will never execute as it is protected
996 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
997 // and later determined that part of the unrolled loop was dead.
998 loop_head->set_trip_count(loop_head->trip_count() / 2);
1000 // Double the count of original iterations in the unrolled loop body.
1001 loop_head->double_unrolled_count();
1003 // -----------
1004 // Step 2: Cut back the trip counter for an unroll amount of 2.
1005 // Loop will normally trip (limit - init)/stride_con. Since it's a
1006 // CountedLoop this is exact (stride divides limit-init exactly).
1007 // We are going to double the loop body, so we want to knock off any
1008 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
1009 Node *span = new (C, 3) SubINode( limit, init );
1010 register_new_node( span, ctrl );
1011 Node *trip = new (C, 3) DivINode( 0, span, stride );
1012 register_new_node( trip, ctrl );
1013 Node *mtwo = _igvn.intcon(-2);
1014 set_ctrl(mtwo, C->root());
1015 Node *rond = new (C, 3) AndINode( trip, mtwo );
1016 register_new_node( rond, ctrl );
1017 Node *spn2 = new (C, 3) MulINode( rond, stride );
1018 register_new_node( spn2, ctrl );
1019 Node *lim2 = new (C, 3) AddINode( spn2, init );
1020 register_new_node( lim2, ctrl );
1022 // Hammer in the new limit
1023 Node *ctrl2 = loop_end->in(0);
1024 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
1025 register_new_node( cmp2, ctrl2 );
1026 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
1027 register_new_node( bol2, ctrl2 );
1028 _igvn.hash_delete(loop_end);
1029 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
1031 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
1032 // Make it a 1-trip test (means at least 2 trips).
1033 if( adjust_min_trip ) {
1034 // Guard test uses an 'opaque' node which is not shared. Hence I
1035 // can edit it's inputs directly. Hammer in the new limit for the
1036 // minimum-trip guard.
1037 assert( opaq->outcnt() == 1, "" );
1038 _igvn.hash_delete(opaq);
1039 opaq->set_req(1, lim2);
1040 }
1042 // ---------
1043 // Step 4: Clone the loop body. Move it inside the loop. This loop body
1044 // represents the odd iterations; since the loop trips an even number of
1045 // times its backedge is never taken. Kill the backedge.
1046 uint dd = dom_depth(loop_head);
1047 clone_loop( loop, old_new, dd );
1049 // Make backedges of the clone equal to backedges of the original.
1050 // Make the fall-in from the original come from the fall-out of the clone.
1051 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
1052 Node* phi = loop_head->fast_out(j);
1053 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
1054 Node *newphi = old_new[phi->_idx];
1055 _igvn.hash_delete( phi );
1056 _igvn.hash_delete( newphi );
1058 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
1059 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
1060 phi ->set_req(LoopNode::LoopBackControl, C->top());
1061 }
1062 }
1063 Node *clone_head = old_new[loop_head->_idx];
1064 _igvn.hash_delete( clone_head );
1065 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
1066 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
1067 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
1068 loop->_head = clone_head; // New loop header
1070 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
1071 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
1073 // Kill the clone's backedge
1074 Node *newcle = old_new[loop_end->_idx];
1075 _igvn.hash_delete( newcle );
1076 Node *one = _igvn.intcon(1);
1077 set_ctrl(one, C->root());
1078 newcle->set_req(1, one);
1079 // Force clone into same loop body
1080 uint max = loop->_body.size();
1081 for( uint k = 0; k < max; k++ ) {
1082 Node *old = loop->_body.at(k);
1083 Node *nnn = old_new[old->_idx];
1084 loop->_body.push(nnn);
1085 if (!has_ctrl(old))
1086 set_loop(nnn, loop);
1087 }
1089 loop->record_for_igvn();
1090 }
1092 //------------------------------do_maximally_unroll----------------------------
1094 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
1095 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1096 assert(cl->trip_count() > 0, "");
1097 #ifndef PRODUCT
1098 if (TraceLoopOpts) {
1099 tty->print("MaxUnroll %d ", cl->trip_count());
1100 loop->dump_head();
1101 }
1102 #endif
1104 // If loop is tripping an odd number of times, peel odd iteration
1105 if ((cl->trip_count() & 1) == 1) {
1106 do_peeling(loop, old_new);
1107 }
1109 // Now its tripping an even number of times remaining. Double loop body.
1110 // Do not adjust pre-guards; they are not needed and do not exist.
1111 if (cl->trip_count() > 0) {
1112 do_unroll(loop, old_new, false);
1113 }
1114 }
1116 //------------------------------dominates_backedge---------------------------------
1117 // Returns true if ctrl is executed on every complete iteration
1118 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
1119 assert(ctrl->is_CFG(), "must be control");
1120 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
1121 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
1122 }
1124 //------------------------------add_constraint---------------------------------
1125 // Constrain the main loop iterations so the condition:
1126 // scale_con * I + offset < limit
1127 // always holds true. That is, either increase the number of iterations in
1128 // the pre-loop or the post-loop until the condition holds true in the main
1129 // loop. Stride, scale, offset and limit are all loop invariant. Further,
1130 // stride and scale are constants (offset and limit often are).
1131 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
1133 // Compute "I :: (limit-offset)/scale_con"
1134 Node *con = new (C, 3) SubINode( limit, offset );
1135 register_new_node( con, pre_ctrl );
1136 Node *scale = _igvn.intcon(scale_con);
1137 set_ctrl(scale, C->root());
1138 Node *X = new (C, 3) DivINode( 0, con, scale );
1139 register_new_node( X, pre_ctrl );
1141 // For positive stride, the pre-loop limit always uses a MAX function
1142 // and the main loop a MIN function. For negative stride these are
1143 // reversed.
1145 // Also for positive stride*scale the affine function is increasing, so the
1146 // pre-loop must check for underflow and the post-loop for overflow.
1147 // Negative stride*scale reverses this; pre-loop checks for overflow and
1148 // post-loop for underflow.
1149 if( stride_con*scale_con > 0 ) {
1150 // Compute I < (limit-offset)/scale_con
1151 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
1152 *main_limit = (stride_con > 0)
1153 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
1154 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
1155 register_new_node( *main_limit, pre_ctrl );
1157 } else {
1158 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
1159 // Add the negation of the main-loop constraint to the pre-loop.
1160 // See footnote [++] below for a derivation of the limit expression.
1161 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
1162 set_ctrl(incr, C->root());
1163 Node *adj = new (C, 3) AddINode( X, incr );
1164 register_new_node( adj, pre_ctrl );
1165 *pre_limit = (scale_con > 0)
1166 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
1167 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
1168 register_new_node( *pre_limit, pre_ctrl );
1170 // [++] Here's the algebra that justifies the pre-loop limit expression:
1171 //
1172 // NOT( scale_con * I + offset < limit )
1173 // ==
1174 // scale_con * I + offset >= limit
1175 // ==
1176 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
1177 // ==
1178 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
1179 // ==
1180 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
1181 // ==
1182 // ( if (scale_con > 0) /*common case*/
1183 // (limit-offset)/scale_con - 1 < I
1184 // else
1185 // (limit-offset)/scale_con + 1 > I
1186 // )
1187 // ( if (scale_con > 0) /*common case*/
1188 // (limit-offset)/scale_con + SGN(-scale_con) < I
1189 // else
1190 // (limit-offset)/scale_con + SGN(-scale_con) > I
1191 }
1192 }
1195 //------------------------------is_scaled_iv---------------------------------
1196 // Return true if exp is a constant times an induction var
1197 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
1198 if (exp == iv) {
1199 if (p_scale != NULL) {
1200 *p_scale = 1;
1201 }
1202 return true;
1203 }
1204 int opc = exp->Opcode();
1205 if (opc == Op_MulI) {
1206 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1207 if (p_scale != NULL) {
1208 *p_scale = exp->in(2)->get_int();
1209 }
1210 return true;
1211 }
1212 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
1213 if (p_scale != NULL) {
1214 *p_scale = exp->in(1)->get_int();
1215 }
1216 return true;
1217 }
1218 } else if (opc == Op_LShiftI) {
1219 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
1220 if (p_scale != NULL) {
1221 *p_scale = 1 << exp->in(2)->get_int();
1222 }
1223 return true;
1224 }
1225 }
1226 return false;
1227 }
1229 //-----------------------------is_scaled_iv_plus_offset------------------------------
1230 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
1231 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
1232 if (is_scaled_iv(exp, iv, p_scale)) {
1233 if (p_offset != NULL) {
1234 Node *zero = _igvn.intcon(0);
1235 set_ctrl(zero, C->root());
1236 *p_offset = zero;
1237 }
1238 return true;
1239 }
1240 int opc = exp->Opcode();
1241 if (opc == Op_AddI) {
1242 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1243 if (p_offset != NULL) {
1244 *p_offset = exp->in(2);
1245 }
1246 return true;
1247 }
1248 if (exp->in(2)->is_Con()) {
1249 Node* offset2 = NULL;
1250 if (depth < 2 &&
1251 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
1252 p_offset != NULL ? &offset2 : NULL, depth+1)) {
1253 if (p_offset != NULL) {
1254 Node *ctrl_off2 = get_ctrl(offset2);
1255 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
1256 register_new_node(offset, ctrl_off2);
1257 *p_offset = offset;
1258 }
1259 return true;
1260 }
1261 }
1262 } else if (opc == Op_SubI) {
1263 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
1264 if (p_offset != NULL) {
1265 Node *zero = _igvn.intcon(0);
1266 set_ctrl(zero, C->root());
1267 Node *ctrl_off = get_ctrl(exp->in(2));
1268 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
1269 register_new_node(offset, ctrl_off);
1270 *p_offset = offset;
1271 }
1272 return true;
1273 }
1274 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
1275 if (p_offset != NULL) {
1276 *p_scale *= -1;
1277 *p_offset = exp->in(1);
1278 }
1279 return true;
1280 }
1281 }
1282 return false;
1283 }
1285 //------------------------------do_range_check---------------------------------
1286 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1287 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
1288 #ifndef PRODUCT
1289 if (PrintOpto && VerifyLoopOptimizations) {
1290 tty->print("Range Check Elimination ");
1291 loop->dump_head();
1292 } else if (TraceLoopOpts) {
1293 tty->print("RangeCheck ");
1294 loop->dump_head();
1295 }
1296 #endif
1297 assert(RangeCheckElimination, "");
1298 CountedLoopNode *cl = loop->_head->as_CountedLoop();
1299 assert(cl->is_main_loop(), "");
1301 // protect against stride not being a constant
1302 if (!cl->stride_is_con())
1303 return;
1305 // Find the trip counter; we are iteration splitting based on it
1306 Node *trip_counter = cl->phi();
1307 // Find the main loop limit; we will trim it's iterations
1308 // to not ever trip end tests
1309 Node *main_limit = cl->limit();
1311 // Need to find the main-loop zero-trip guard
1312 Node *ctrl = cl->in(LoopNode::EntryControl);
1313 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
1314 Node *iffm = ctrl->in(0);
1315 assert(iffm->Opcode() == Op_If, "");
1316 Node *bolzm = iffm->in(1);
1317 assert(bolzm->Opcode() == Op_Bool, "");
1318 Node *cmpzm = bolzm->in(1);
1319 assert(cmpzm->is_Cmp(), "");
1320 Node *opqzm = cmpzm->in(2);
1321 // Can not optimize a loop if pre-loop Opaque1 node is optimized
1322 // away and then another round of loop opts attempted.
1323 if (opqzm->Opcode() != Op_Opaque1)
1324 return;
1325 assert(opqzm->in(1) == main_limit, "do not understand situation");
1327 // Find the pre-loop limit; we will expand it's iterations to
1328 // not ever trip low tests.
1329 Node *p_f = iffm->in(0);
1330 assert(p_f->Opcode() == Op_IfFalse, "");
1331 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
1332 assert(pre_end->loopnode()->is_pre_loop(), "");
1333 Node *pre_opaq1 = pre_end->limit();
1334 // Occasionally it's possible for a pre-loop Opaque1 node to be
1335 // optimized away and then another round of loop opts attempted.
1336 // We can not optimize this particular loop in that case.
1337 if (pre_opaq1->Opcode() != Op_Opaque1)
1338 return;
1339 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
1340 Node *pre_limit = pre_opaq->in(1);
1342 // Where do we put new limit calculations
1343 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
1345 // Ensure the original loop limit is available from the
1346 // pre-loop Opaque1 node.
1347 Node *orig_limit = pre_opaq->original_loop_limit();
1348 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
1349 return;
1351 // Must know if its a count-up or count-down loop
1353 int stride_con = cl->stride_con();
1354 Node *zero = _igvn.intcon(0);
1355 Node *one = _igvn.intcon(1);
1356 set_ctrl(zero, C->root());
1357 set_ctrl(one, C->root());
1359 // Range checks that do not dominate the loop backedge (ie.
1360 // conditionally executed) can lengthen the pre loop limit beyond
1361 // the original loop limit. To prevent this, the pre limit is
1362 // (for stride > 0) MINed with the original loop limit (MAXed
1363 // stride < 0) when some range_check (rc) is conditionally
1364 // executed.
1365 bool conditional_rc = false;
1367 // Check loop body for tests of trip-counter plus loop-invariant vs
1368 // loop-invariant.
1369 for( uint i = 0; i < loop->_body.size(); i++ ) {
1370 Node *iff = loop->_body[i];
1371 if( iff->Opcode() == Op_If ) { // Test?
1373 // Test is an IfNode, has 2 projections. If BOTH are in the loop
1374 // we need loop unswitching instead of iteration splitting.
1375 Node *exit = loop->is_loop_exit(iff);
1376 if( !exit ) continue;
1377 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
1379 // Get boolean condition to test
1380 Node *i1 = iff->in(1);
1381 if( !i1->is_Bool() ) continue;
1382 BoolNode *bol = i1->as_Bool();
1383 BoolTest b_test = bol->_test;
1384 // Flip sense of test if exit condition is flipped
1385 if( flip )
1386 b_test = b_test.negate();
1388 // Get compare
1389 Node *cmp = bol->in(1);
1391 // Look for trip_counter + offset vs limit
1392 Node *rc_exp = cmp->in(1);
1393 Node *limit = cmp->in(2);
1394 jint scale_con= 1; // Assume trip counter not scaled
1396 Node *limit_c = get_ctrl(limit);
1397 if( loop->is_member(get_loop(limit_c) ) ) {
1398 // Compare might have operands swapped; commute them
1399 b_test = b_test.commute();
1400 rc_exp = cmp->in(2);
1401 limit = cmp->in(1);
1402 limit_c = get_ctrl(limit);
1403 if( loop->is_member(get_loop(limit_c) ) )
1404 continue; // Both inputs are loop varying; cannot RCE
1405 }
1406 // Here we know 'limit' is loop invariant
1408 // 'limit' maybe pinned below the zero trip test (probably from a
1409 // previous round of rce), in which case, it can't be used in the
1410 // zero trip test expression which must occur before the zero test's if.
1411 if( limit_c == ctrl ) {
1412 continue; // Don't rce this check but continue looking for other candidates.
1413 }
1415 // Check for scaled induction variable plus an offset
1416 Node *offset = NULL;
1418 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
1419 continue;
1420 }
1422 Node *offset_c = get_ctrl(offset);
1423 if( loop->is_member( get_loop(offset_c) ) )
1424 continue; // Offset is not really loop invariant
1425 // Here we know 'offset' is loop invariant.
1427 // As above for the 'limit', the 'offset' maybe pinned below the
1428 // zero trip test.
1429 if( offset_c == ctrl ) {
1430 continue; // Don't rce this check but continue looking for other candidates.
1431 }
1433 // At this point we have the expression as:
1434 // scale_con * trip_counter + offset :: limit
1435 // where scale_con, offset and limit are loop invariant. Trip_counter
1436 // monotonically increases by stride_con, a constant. Both (or either)
1437 // stride_con and scale_con can be negative which will flip about the
1438 // sense of the test.
1440 // Adjust pre and main loop limits to guard the correct iteration set
1441 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
1442 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
1443 // The overflow limit: scale*I+offset < limit
1444 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1445 // The underflow limit: 0 <= scale*I+offset.
1446 // Some math yields: -scale*I-(offset+1) < 0
1447 Node *plus_one = new (C, 3) AddINode( offset, one );
1448 register_new_node( plus_one, pre_ctrl );
1449 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
1450 register_new_node( neg_offset, pre_ctrl );
1451 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
1452 if (!conditional_rc) {
1453 conditional_rc = !loop->dominates_backedge(iff);
1454 }
1455 } else {
1456 #ifndef PRODUCT
1457 if( PrintOpto )
1458 tty->print_cr("missed RCE opportunity");
1459 #endif
1460 continue; // In release mode, ignore it
1461 }
1462 } else { // Otherwise work on normal compares
1463 switch( b_test._test ) {
1464 case BoolTest::ge: // Convert X >= Y to -X <= -Y
1465 scale_con = -scale_con;
1466 offset = new (C, 3) SubINode( zero, offset );
1467 register_new_node( offset, pre_ctrl );
1468 limit = new (C, 3) SubINode( zero, limit );
1469 register_new_node( limit, pre_ctrl );
1470 // Fall into LE case
1471 case BoolTest::le: // Convert X <= Y to X < Y+1
1472 limit = new (C, 3) AddINode( limit, one );
1473 register_new_node( limit, pre_ctrl );
1474 // Fall into LT case
1475 case BoolTest::lt:
1476 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
1477 if (!conditional_rc) {
1478 conditional_rc = !loop->dominates_backedge(iff);
1479 }
1480 break;
1481 default:
1482 #ifndef PRODUCT
1483 if( PrintOpto )
1484 tty->print_cr("missed RCE opportunity");
1485 #endif
1486 continue; // Unhandled case
1487 }
1488 }
1490 // Kill the eliminated test
1491 C->set_major_progress();
1492 Node *kill_con = _igvn.intcon( 1-flip );
1493 set_ctrl(kill_con, C->root());
1494 _igvn.hash_delete(iff);
1495 iff->set_req(1, kill_con);
1496 _igvn._worklist.push(iff);
1497 // Find surviving projection
1498 assert(iff->is_If(), "");
1499 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
1500 // Find loads off the surviving projection; remove their control edge
1501 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
1502 Node* cd = dp->fast_out(i); // Control-dependent node
1503 if( cd->is_Load() ) { // Loads can now float around in the loop
1504 _igvn.hash_delete(cd);
1505 // Allow the load to float around in the loop, or before it
1506 // but NOT before the pre-loop.
1507 cd->set_req(0, ctrl); // ctrl, not NULL
1508 _igvn._worklist.push(cd);
1509 --i;
1510 --imax;
1511 }
1512 }
1514 } // End of is IF
1516 }
1518 // Update loop limits
1519 if (conditional_rc) {
1520 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
1521 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
1522 register_new_node(pre_limit, pre_ctrl);
1523 }
1524 _igvn.hash_delete(pre_opaq);
1525 pre_opaq->set_req(1, pre_limit);
1527 // Note:: we are making the main loop limit no longer precise;
1528 // need to round up based on stride.
1529 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
1530 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
1531 // Hopefully, compiler will optimize for powers of 2.
1532 Node *ctrl = get_ctrl(main_limit);
1533 Node *stride = cl->stride();
1534 Node *init = cl->init_trip();
1535 Node *span = new (C, 3) SubINode(main_limit,init);
1536 register_new_node(span,ctrl);
1537 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
1538 Node *add = new (C, 3) AddINode(span,rndup);
1539 register_new_node(add,ctrl);
1540 Node *div = new (C, 3) DivINode(0,add,stride);
1541 register_new_node(div,ctrl);
1542 Node *mul = new (C, 3) MulINode(div,stride);
1543 register_new_node(mul,ctrl);
1544 Node *newlim = new (C, 3) AddINode(mul,init);
1545 register_new_node(newlim,ctrl);
1546 main_limit = newlim;
1547 }
1549 Node *main_cle = cl->loopexit();
1550 Node *main_bol = main_cle->in(1);
1551 // Hacking loop bounds; need private copies of exit test
1552 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
1553 _igvn.hash_delete(main_cle);
1554 main_bol = main_bol->clone();// Clone a private BoolNode
1555 register_new_node( main_bol, main_cle->in(0) );
1556 main_cle->set_req(1,main_bol);
1557 }
1558 Node *main_cmp = main_bol->in(1);
1559 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
1560 _igvn.hash_delete(main_bol);
1561 main_cmp = main_cmp->clone();// Clone a private CmpNode
1562 register_new_node( main_cmp, main_cle->in(0) );
1563 main_bol->set_req(1,main_cmp);
1564 }
1565 // Hack the now-private loop bounds
1566 _igvn.hash_delete(main_cmp);
1567 main_cmp->set_req(2, main_limit);
1568 _igvn._worklist.push(main_cmp);
1569 // The OpaqueNode is unshared by design
1570 _igvn.hash_delete(opqzm);
1571 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
1572 opqzm->set_req(1,main_limit);
1573 _igvn._worklist.push(opqzm);
1574 }
1576 //------------------------------DCE_loop_body----------------------------------
1577 // Remove simplistic dead code from loop body
1578 void IdealLoopTree::DCE_loop_body() {
1579 for( uint i = 0; i < _body.size(); i++ )
1580 if( _body.at(i)->outcnt() == 0 )
1581 _body.map( i--, _body.pop() );
1582 }
1585 //------------------------------adjust_loop_exit_prob--------------------------
1586 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
1587 // Replace with a 1-in-10 exit guess.
1588 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
1589 Node *test = tail();
1590 while( test != _head ) {
1591 uint top = test->Opcode();
1592 if( top == Op_IfTrue || top == Op_IfFalse ) {
1593 int test_con = ((ProjNode*)test)->_con;
1594 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
1595 IfNode *iff = test->in(0)->as_If();
1596 if( iff->outcnt() == 2 ) { // Ignore dead tests
1597 Node *bol = iff->in(1);
1598 if( bol && bol->req() > 1 && bol->in(1) &&
1599 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
1600 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
1601 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
1602 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
1603 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
1604 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
1605 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
1606 return; // Allocation loops RARELY take backedge
1607 // Find the OTHER exit path from the IF
1608 Node* ex = iff->proj_out(1-test_con);
1609 float p = iff->_prob;
1610 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
1611 if( top == Op_IfTrue ) {
1612 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
1613 iff->_prob = PROB_STATIC_FREQUENT;
1614 }
1615 } else {
1616 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
1617 iff->_prob = PROB_STATIC_INFREQUENT;
1618 }
1619 }
1620 }
1621 }
1622 }
1623 test = phase->idom(test);
1624 }
1625 }
1628 //------------------------------policy_do_remove_empty_loop--------------------
1629 // Micro-benchmark spamming. Policy is to always remove empty loops.
1630 // The 'DO' part is to replace the trip counter with the value it will
1631 // have on the last iteration. This will break the loop.
1632 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
1633 // Minimum size must be empty loop
1634 if (_body.size() > 7/*number of nodes in an empty loop*/)
1635 return false;
1637 if (!_head->is_CountedLoop())
1638 return false; // Dead loop
1639 CountedLoopNode *cl = _head->as_CountedLoop();
1640 if (!cl->loopexit())
1641 return false; // Malformed loop
1642 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
1643 return false; // Infinite loop
1645 #ifdef ASSERT
1646 // Ensure only one phi which is the iv.
1647 Node* iv = NULL;
1648 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
1649 Node* n = cl->fast_out(i);
1650 if (n->Opcode() == Op_Phi) {
1651 assert(iv == NULL, "Too many phis" );
1652 iv = n;
1653 }
1654 }
1655 assert(iv == cl->phi(), "Wrong phi" );
1656 #endif
1658 // main and post loops have explicitly created zero trip guard
1659 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
1660 if (needs_guard) {
1661 // Check for an obvious zero trip guard.
1662 Node* inctrl = cl->in(LoopNode::EntryControl);
1663 if (inctrl->Opcode() == Op_IfTrue) {
1664 // The test should look like just the backedge of a CountedLoop
1665 Node* iff = inctrl->in(0);
1666 if (iff->is_If()) {
1667 Node* bol = iff->in(1);
1668 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
1669 Node* cmp = bol->in(1);
1670 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
1671 needs_guard = false;
1672 }
1673 }
1674 }
1675 }
1676 }
1678 #ifndef PRODUCT
1679 if (PrintOpto) {
1680 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
1681 this->dump_head();
1682 } else if (TraceLoopOpts) {
1683 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
1684 this->dump_head();
1685 }
1686 #endif
1688 if (needs_guard) {
1689 // Peel the loop to ensure there's a zero trip guard
1690 Node_List old_new;
1691 phase->do_peeling(this, old_new);
1692 }
1694 // Replace the phi at loop head with the final value of the last
1695 // iteration. Then the CountedLoopEnd will collapse (backedge never
1696 // taken) and all loop-invariant uses of the exit values will be correct.
1697 Node *phi = cl->phi();
1698 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
1699 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
1700 phase->_igvn.replace_node(phi,final);
1701 phase->C->set_major_progress();
1702 return true;
1703 }
1706 //=============================================================================
1707 //------------------------------iteration_split_impl---------------------------
1708 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
1709 // Check and remove empty loops (spam micro-benchmarks)
1710 if( policy_do_remove_empty_loop(phase) )
1711 return true; // Here we removed an empty loop
1713 bool should_peel = policy_peeling(phase); // Should we peel?
1715 bool should_unswitch = policy_unswitching(phase);
1717 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
1718 // This removes loop-invariant tests (usually null checks).
1719 if( !_head->is_CountedLoop() ) { // Non-counted loop
1720 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
1721 // Partial peel succeeded so terminate this round of loop opts
1722 return false;
1723 }
1724 if( should_peel ) { // Should we peel?
1725 #ifndef PRODUCT
1726 if (PrintOpto) tty->print_cr("should_peel");
1727 #endif
1728 phase->do_peeling(this,old_new);
1729 } else if( should_unswitch ) {
1730 phase->do_unswitching(this, old_new);
1731 }
1732 return true;
1733 }
1734 CountedLoopNode *cl = _head->as_CountedLoop();
1736 if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops
1738 // Do nothing special to pre- and post- loops
1739 if( cl->is_pre_loop() || cl->is_post_loop() ) return true;
1741 // Compute loop trip count from profile data
1742 compute_profile_trip_cnt(phase);
1744 // Before attempting fancy unrolling, RCE or alignment, see if we want
1745 // to completely unroll this loop or do loop unswitching.
1746 if( cl->is_normal_loop() ) {
1747 if (should_unswitch) {
1748 phase->do_unswitching(this, old_new);
1749 return true;
1750 }
1751 bool should_maximally_unroll = policy_maximally_unroll(phase);
1752 if( should_maximally_unroll ) {
1753 // Here we did some unrolling and peeling. Eventually we will
1754 // completely unroll this loop and it will no longer be a loop.
1755 phase->do_maximally_unroll(this,old_new);
1756 return true;
1757 }
1758 }
1761 // Counted loops may be peeled, may need some iterations run up
1762 // front for RCE, and may want to align loop refs to a cache
1763 // line. Thus we clone a full loop up front whose trip count is
1764 // at least 1 (if peeling), but may be several more.
1766 // The main loop will start cache-line aligned with at least 1
1767 // iteration of the unrolled body (zero-trip test required) and
1768 // will have some range checks removed.
1770 // A post-loop will finish any odd iterations (leftover after
1771 // unrolling), plus any needed for RCE purposes.
1773 bool should_unroll = policy_unroll(phase);
1775 bool should_rce = policy_range_check(phase);
1777 bool should_align = policy_align(phase);
1779 // If not RCE'ing (iteration splitting) or Aligning, then we do not
1780 // need a pre-loop. We may still need to peel an initial iteration but
1781 // we will not be needing an unknown number of pre-iterations.
1782 //
1783 // Basically, if may_rce_align reports FALSE first time through,
1784 // we will not be able to later do RCE or Aligning on this loop.
1785 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
1787 // If we have any of these conditions (RCE, alignment, unrolling) met, then
1788 // we switch to the pre-/main-/post-loop model. This model also covers
1789 // peeling.
1790 if( should_rce || should_align || should_unroll ) {
1791 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
1792 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
1794 // Adjust the pre- and main-loop limits to let the pre and post loops run
1795 // with full checks, but the main-loop with no checks. Remove said
1796 // checks from the main body.
1797 if( should_rce )
1798 phase->do_range_check(this,old_new);
1800 // Double loop body for unrolling. Adjust the minimum-trip test (will do
1801 // twice as many iterations as before) and the main body limit (only do
1802 // an even number of trips). If we are peeling, we might enable some RCE
1803 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
1804 // peeling.
1805 if( should_unroll && !should_peel )
1806 phase->do_unroll(this,old_new, true);
1808 // Adjust the pre-loop limits to align the main body
1809 // iterations.
1810 if( should_align )
1811 Unimplemented();
1813 } else { // Else we have an unchanged counted loop
1814 if( should_peel ) // Might want to peel but do nothing else
1815 phase->do_peeling(this,old_new);
1816 }
1817 return true;
1818 }
1821 //=============================================================================
1822 //------------------------------iteration_split--------------------------------
1823 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
1824 // Recursively iteration split nested loops
1825 if (_child && !_child->iteration_split(phase, old_new))
1826 return false;
1828 // Clean out prior deadwood
1829 DCE_loop_body();
1832 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1833 // Replace with a 1-in-10 exit guess.
1834 if (_parent /*not the root loop*/ &&
1835 !_irreducible &&
1836 // Also ignore the occasional dead backedge
1837 !tail()->is_top()) {
1838 adjust_loop_exit_prob(phase);
1839 }
1841 // Gate unrolling, RCE and peeling efforts.
1842 if (!_child && // If not an inner loop, do not split
1843 !_irreducible &&
1844 _allow_optimizations &&
1845 !tail()->is_top()) { // Also ignore the occasional dead backedge
1846 if (!_has_call) {
1847 if (!iteration_split_impl(phase, old_new)) {
1848 return false;
1849 }
1850 } else if (policy_unswitching(phase)) {
1851 phase->do_unswitching(this, old_new);
1852 }
1853 }
1855 // Minor offset re-organization to remove loop-fallout uses of
1856 // trip counter when there was no major reshaping.
1857 phase->reorg_offsets(this);
1859 if (_next && !_next->iteration_split(phase, old_new))
1860 return false;
1861 return true;
1862 }
1864 //-------------------------------is_uncommon_trap_proj----------------------------
1865 // Return true if proj is the form of "proj->[region->..]call_uct"
1866 bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) {
1867 int path_limit = 10;
1868 assert(proj, "invalid argument");
1869 Node* out = proj;
1870 for (int ct = 0; ct < path_limit; ct++) {
1871 out = out->unique_ctrl_out();
1872 if (out == NULL || out->is_Root() || out->is_Start())
1873 return false;
1874 if (out->is_CallStaticJava()) {
1875 int req = out->as_CallStaticJava()->uncommon_trap_request();
1876 if (req != 0) {
1877 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
1878 if (trap_reason == reason || reason == Deoptimization::Reason_none) {
1879 return true;
1880 }
1881 }
1882 return false; // don't do further after call
1883 }
1884 }
1885 return false;
1886 }
1888 //-------------------------------is_uncommon_trap_if_pattern-------------------------
1889 // Return true for "if(test)-> proj -> ...
1890 // |
1891 // V
1892 // other_proj->[region->..]call_uct"
1893 //
1894 // "must_reason_predicate" means the uct reason must be Reason_predicate
1895 bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) {
1896 Node *in0 = proj->in(0);
1897 if (!in0->is_If()) return false;
1898 // Variation of a dead If node.
1899 if (in0->outcnt() < 2) return false;
1900 IfNode* iff = in0->as_If();
1902 // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
1903 if (reason != Deoptimization::Reason_none) {
1904 if (iff->in(1)->Opcode() != Op_Conv2B ||
1905 iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
1906 return false;
1907 }
1908 }
1910 ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
1911 return is_uncommon_trap_proj(other_proj, reason);
1912 }
1914 //-------------------------------register_control-------------------------
1915 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
1916 assert(n->is_CFG(), "must be control node");
1917 _igvn.register_new_node_with_optimizer(n);
1918 loop->_body.push(n);
1919 set_loop(n, loop);
1920 // When called from beautify_loops() idom is not constructed yet.
1921 if (_idom != NULL) {
1922 set_idom(n, pred, dom_depth(pred));
1923 }
1924 }
1926 //------------------------------create_new_if_for_predicate------------------------
1927 // create a new if above the uct_if_pattern for the predicate to be promoted.
1928 //
1929 // before after
1930 // ---------- ----------
1931 // ctrl ctrl
1932 // | |
1933 // | |
1934 // v v
1935 // iff new_iff
1936 // / \ / \
1937 // / \ / \
1938 // v v v v
1939 // uncommon_proj cont_proj if_uct if_cont
1940 // \ | | | |
1941 // \ | | | |
1942 // v v v | v
1943 // rgn loop | iff
1944 // | | / \
1945 // | | / \
1946 // v | v v
1947 // uncommon_trap | uncommon_proj cont_proj
1948 // \ \ | |
1949 // \ \ | |
1950 // v v v v
1951 // rgn loop
1952 // |
1953 // |
1954 // v
1955 // uncommon_trap
1956 //
1957 //
1958 // We will create a region to guard the uct call if there is no one there.
1959 // The true projecttion (if_cont) of the new_iff is returned.
1960 // This code is also used to clone predicates to clonned loops.
1961 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
1962 Deoptimization::DeoptReason reason) {
1963 assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
1964 IfNode* iff = cont_proj->in(0)->as_If();
1966 ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
1967 Node *rgn = uncommon_proj->unique_ctrl_out();
1968 assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct");
1970 if (!rgn->is_Region()) { // create a region to guard the call
1971 assert(rgn->is_Call(), "must be call uct");
1972 CallNode* call = rgn->as_Call();
1973 IdealLoopTree* loop = get_loop(call);
1974 rgn = new (C, 1) RegionNode(1);
1975 rgn->add_req(uncommon_proj);
1976 register_control(rgn, loop, uncommon_proj);
1977 _igvn.hash_delete(call);
1978 call->set_req(0, rgn);
1979 // When called from beautify_loops() idom is not constructed yet.
1980 if (_idom != NULL) {
1981 set_idom(call, rgn, dom_depth(rgn));
1982 }
1983 }
1985 Node* entry = iff->in(0);
1986 if (new_entry != NULL) {
1987 // Clonning the predicate to new location.
1988 entry = new_entry;
1989 }
1990 // Create new_iff
1991 IdealLoopTree* lp = get_loop(entry);
1992 IfNode *new_iff = new (C, 2) IfNode(entry, NULL, iff->_prob, iff->_fcnt);
1993 register_control(new_iff, lp, entry);
1994 Node *if_cont = new (C, 1) IfTrueNode(new_iff);
1995 Node *if_uct = new (C, 1) IfFalseNode(new_iff);
1996 if (cont_proj->is_IfFalse()) {
1997 // Swap
1998 Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp;
1999 }
2000 register_control(if_cont, lp, new_iff);
2001 register_control(if_uct, get_loop(rgn), new_iff);
2003 // if_uct to rgn
2004 _igvn.hash_delete(rgn);
2005 rgn->add_req(if_uct);
2006 // When called from beautify_loops() idom is not constructed yet.
2007 if (_idom != NULL) {
2008 Node* ridom = idom(rgn);
2009 Node* nrdom = dom_lca(ridom, new_iff);
2010 set_idom(rgn, nrdom, dom_depth(rgn));
2011 }
2012 // rgn must have no phis
2013 assert(!rgn->as_Region()->has_phi(), "region must have no phis");
2015 if (new_entry == NULL) {
2016 // Attach if_cont to iff
2017 _igvn.hash_delete(iff);
2018 iff->set_req(0, if_cont);
2019 if (_idom != NULL) {
2020 set_idom(iff, if_cont, dom_depth(iff));
2021 }
2022 }
2023 return if_cont->as_Proj();
2024 }
2026 //--------------------------find_predicate_insertion_point-------------------
2027 // Find a good location to insert a predicate
2028 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
2029 if (start_c == NULL || !start_c->is_Proj())
2030 return NULL;
2031 if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
2032 return start_c->as_Proj();
2033 }
2034 return NULL;
2035 }
2037 //--------------------------find_predicate------------------------------------
2038 // Find a predicate
2039 Node* PhaseIdealLoop::find_predicate(Node* entry) {
2040 Node* predicate = NULL;
2041 if (UseLoopPredicate) {
2042 predicate = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2043 if (predicate != NULL) { // right pattern that can be used by loop predication
2044 assert(entry->in(0)->in(1)->in(1)->Opcode()==Op_Opaque1, "must be");
2045 return entry;
2046 }
2047 }
2048 return NULL;
2049 }
2051 //------------------------------Invariance-----------------------------------
2052 // Helper class for loop_predication_impl to compute invariance on the fly and
2053 // clone invariants.
2054 class Invariance : public StackObj {
2055 VectorSet _visited, _invariant;
2056 Node_Stack _stack;
2057 VectorSet _clone_visited;
2058 Node_List _old_new; // map of old to new (clone)
2059 IdealLoopTree* _lpt;
2060 PhaseIdealLoop* _phase;
2062 // Helper function to set up the invariance for invariance computation
2063 // If n is a known invariant, set up directly. Otherwise, look up the
2064 // the possibility to push n onto the stack for further processing.
2065 void visit(Node* use, Node* n) {
2066 if (_lpt->is_invariant(n)) { // known invariant
2067 _invariant.set(n->_idx);
2068 } else if (!n->is_CFG()) {
2069 Node *n_ctrl = _phase->ctrl_or_self(n);
2070 Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG
2071 if (_phase->is_dominator(n_ctrl, u_ctrl)) {
2072 _stack.push(n, n->in(0) == NULL ? 1 : 0);
2073 }
2074 }
2075 }
2077 // Compute invariance for "the_node" and (possibly) all its inputs recursively
2078 // on the fly
2079 void compute_invariance(Node* n) {
2080 assert(_visited.test(n->_idx), "must be");
2081 visit(n, n);
2082 while (_stack.is_nonempty()) {
2083 Node* n = _stack.node();
2084 uint idx = _stack.index();
2085 if (idx == n->req()) { // all inputs are processed
2086 _stack.pop();
2087 // n is invariant if it's inputs are all invariant
2088 bool all_inputs_invariant = true;
2089 for (uint i = 0; i < n->req(); i++) {
2090 Node* in = n->in(i);
2091 if (in == NULL) continue;
2092 assert(_visited.test(in->_idx), "must have visited input");
2093 if (!_invariant.test(in->_idx)) { // bad guy
2094 all_inputs_invariant = false;
2095 break;
2096 }
2097 }
2098 if (all_inputs_invariant) {
2099 _invariant.set(n->_idx); // I am a invariant too
2100 }
2101 } else { // process next input
2102 _stack.set_index(idx + 1);
2103 Node* m = n->in(idx);
2104 if (m != NULL && !_visited.test_set(m->_idx)) {
2105 visit(n, m);
2106 }
2107 }
2108 }
2109 }
2111 // Helper function to set up _old_new map for clone_nodes.
2112 // If n is a known invariant, set up directly ("clone" of n == n).
2113 // Otherwise, push n onto the stack for real cloning.
2114 void clone_visit(Node* n) {
2115 assert(_invariant.test(n->_idx), "must be invariant");
2116 if (_lpt->is_invariant(n)) { // known invariant
2117 _old_new.map(n->_idx, n);
2118 } else{ // to be cloned
2119 assert (!n->is_CFG(), "should not see CFG here");
2120 _stack.push(n, n->in(0) == NULL ? 1 : 0);
2121 }
2122 }
2124 // Clone "n" and (possibly) all its inputs recursively
2125 void clone_nodes(Node* n, Node* ctrl) {
2126 clone_visit(n);
2127 while (_stack.is_nonempty()) {
2128 Node* n = _stack.node();
2129 uint idx = _stack.index();
2130 if (idx == n->req()) { // all inputs processed, clone n!
2131 _stack.pop();
2132 // clone invariant node
2133 Node* n_cl = n->clone();
2134 _old_new.map(n->_idx, n_cl);
2135 _phase->register_new_node(n_cl, ctrl);
2136 for (uint i = 0; i < n->req(); i++) {
2137 Node* in = n_cl->in(i);
2138 if (in == NULL) continue;
2139 n_cl->set_req(i, _old_new[in->_idx]);
2140 }
2141 } else { // process next input
2142 _stack.set_index(idx + 1);
2143 Node* m = n->in(idx);
2144 if (m != NULL && !_clone_visited.test_set(m->_idx)) {
2145 clone_visit(m); // visit the input
2146 }
2147 }
2148 }
2149 }
2151 public:
2152 Invariance(Arena* area, IdealLoopTree* lpt) :
2153 _lpt(lpt), _phase(lpt->_phase),
2154 _visited(area), _invariant(area), _stack(area, 10 /* guess */),
2155 _clone_visited(area), _old_new(area)
2156 {}
2158 // Map old to n for invariance computation and clone
2159 void map_ctrl(Node* old, Node* n) {
2160 assert(old->is_CFG() && n->is_CFG(), "must be");
2161 _old_new.map(old->_idx, n); // "clone" of old is n
2162 _invariant.set(old->_idx); // old is invariant
2163 _clone_visited.set(old->_idx);
2164 }
2166 // Driver function to compute invariance
2167 bool is_invariant(Node* n) {
2168 if (!_visited.test_set(n->_idx))
2169 compute_invariance(n);
2170 return (_invariant.test(n->_idx) != 0);
2171 }
2173 // Driver function to clone invariant
2174 Node* clone(Node* n, Node* ctrl) {
2175 assert(ctrl->is_CFG(), "must be");
2176 assert(_invariant.test(n->_idx), "must be an invariant");
2177 if (!_clone_visited.test(n->_idx))
2178 clone_nodes(n, ctrl);
2179 return _old_new[n->_idx];
2180 }
2181 };
2183 //------------------------------is_range_check_if -----------------------------------
2184 // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format
2185 // Note: this function is particularly designed for loop predication. We require load_range
2186 // and offset to be loop invariant computed on the fly by "invar"
2187 bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const {
2188 if (!is_loop_exit(iff)) {
2189 return false;
2190 }
2191 if (!iff->in(1)->is_Bool()) {
2192 return false;
2193 }
2194 const BoolNode *bol = iff->in(1)->as_Bool();
2195 if (bol->_test._test != BoolTest::lt) {
2196 return false;
2197 }
2198 if (!bol->in(1)->is_Cmp()) {
2199 return false;
2200 }
2201 const CmpNode *cmp = bol->in(1)->as_Cmp();
2202 if (cmp->Opcode() != Op_CmpU ) {
2203 return false;
2204 }
2205 Node* range = cmp->in(2);
2206 if (range->Opcode() != Op_LoadRange) {
2207 const TypeInt* tint = phase->_igvn.type(range)->isa_int();
2208 if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) {
2209 // Allow predication on positive values that aren't LoadRanges.
2210 // This allows optimization of loops where the length of the
2211 // array is a known value and doesn't need to be loaded back
2212 // from the array.
2213 return false;
2214 }
2215 }
2216 if (!invar.is_invariant(range)) {
2217 return false;
2218 }
2219 Node *iv = _head->as_CountedLoop()->phi();
2220 int scale = 0;
2221 Node *offset = NULL;
2222 if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) {
2223 return false;
2224 }
2225 if(offset && !invar.is_invariant(offset)) { // offset must be invariant
2226 return false;
2227 }
2228 return true;
2229 }
2231 //------------------------------rc_predicate-----------------------------------
2232 // Create a range check predicate
2233 //
2234 // for (i = init; i < limit; i += stride) {
2235 // a[scale*i+offset]
2236 // }
2237 //
2238 // Compute max(scale*i + offset) for init <= i < limit and build the predicate
2239 // as "max(scale*i + offset) u< a.length".
2240 //
2241 // There are two cases for max(scale*i + offset):
2242 // (1) stride*scale > 0
2243 // max(scale*i + offset) = scale*(limit-stride) + offset
2244 // (2) stride*scale < 0
2245 // max(scale*i + offset) = scale*init + offset
2246 BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl,
2247 int scale, Node* offset,
2248 Node* init, Node* limit, Node* stride,
2249 Node* range, bool upper) {
2250 DEBUG_ONLY(ttyLocker ttyl);
2251 if (TraceLoopPredicate) tty->print("rc_predicate ");
2253 Node* max_idx_expr = init;
2254 int stride_con = stride->get_int();
2255 if ((stride_con > 0) == (scale > 0) == upper) {
2256 max_idx_expr = new (C, 3) SubINode(limit, stride);
2257 register_new_node(max_idx_expr, ctrl);
2258 if (TraceLoopPredicate) tty->print("(limit - stride) ");
2259 } else {
2260 if (TraceLoopPredicate) tty->print("init ");
2261 }
2263 if (scale != 1) {
2264 ConNode* con_scale = _igvn.intcon(scale);
2265 max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale);
2266 register_new_node(max_idx_expr, ctrl);
2267 if (TraceLoopPredicate) tty->print("* %d ", scale);
2268 }
2270 if (offset && (!offset->is_Con() || offset->get_int() != 0)){
2271 max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset);
2272 register_new_node(max_idx_expr, ctrl);
2273 if (TraceLoopPredicate)
2274 if (offset->is_Con()) tty->print("+ %d ", offset->get_int());
2275 else tty->print("+ offset ");
2276 }
2278 CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range);
2279 register_new_node(cmp, ctrl);
2280 BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt);
2281 register_new_node(bol, ctrl);
2283 if (TraceLoopPredicate) tty->print_cr("<u range");
2284 return bol;
2285 }
2287 //------------------------------ loop_predication_impl--------------------------
2288 // Insert loop predicates for null checks and range checks
2289 bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
2290 if (!UseLoopPredicate) return false;
2292 if (!loop->_head->is_Loop()) {
2293 // Could be a simple region when irreducible loops are present.
2294 return false;
2295 }
2297 if (loop->_head->unique_ctrl_out()->Opcode() == Op_NeverBranch) {
2298 // do nothing for infinite loops
2299 return false;
2300 }
2302 CountedLoopNode *cl = NULL;
2303 if (loop->_head->is_CountedLoop()) {
2304 cl = loop->_head->as_CountedLoop();
2305 // do nothing for iteration-splitted loops
2306 if (!cl->is_normal_loop()) return false;
2307 }
2309 LoopNode *lpn = loop->_head->as_Loop();
2310 Node* entry = lpn->in(LoopNode::EntryControl);
2312 ProjNode *predicate_proj = find_predicate_insertion_point(entry, Deoptimization::Reason_predicate);
2313 if (!predicate_proj) {
2314 #ifndef PRODUCT
2315 if (TraceLoopPredicate) {
2316 tty->print("missing predicate:");
2317 loop->dump_head();
2318 lpn->dump(1);
2319 }
2320 #endif
2321 return false;
2322 }
2323 ConNode* zero = _igvn.intcon(0);
2324 set_ctrl(zero, C->root());
2326 ResourceArea *area = Thread::current()->resource_area();
2327 Invariance invar(area, loop);
2329 // Create list of if-projs such that a newer proj dominates all older
2330 // projs in the list, and they all dominate loop->tail()
2331 Node_List if_proj_list(area);
2332 LoopNode *head = loop->_head->as_Loop();
2333 Node *current_proj = loop->tail(); //start from tail
2334 while ( current_proj != head ) {
2335 if (loop == get_loop(current_proj) && // still in the loop ?
2336 current_proj->is_Proj() && // is a projection ?
2337 current_proj->in(0)->Opcode() == Op_If) { // is a if projection ?
2338 if_proj_list.push(current_proj);
2339 }
2340 current_proj = idom(current_proj);
2341 }
2343 bool hoisted = false; // true if at least one proj is promoted
2344 while (if_proj_list.size() > 0) {
2345 // Following are changed to nonnull when a predicate can be hoisted
2346 ProjNode* new_predicate_proj = NULL;
2348 ProjNode* proj = if_proj_list.pop()->as_Proj();
2349 IfNode* iff = proj->in(0)->as_If();
2351 if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
2352 if (loop->is_loop_exit(iff)) {
2353 // stop processing the remaining projs in the list because the execution of them
2354 // depends on the condition of "iff" (iff->in(1)).
2355 break;
2356 } else {
2357 // Both arms are inside the loop. There are two cases:
2358 // (1) there is one backward branch. In this case, any remaining proj
2359 // in the if_proj list post-dominates "iff". So, the condition of "iff"
2360 // does not determine the execution the remining projs directly, and we
2361 // can safely continue.
2362 // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj"
2363 // does not dominate loop->tail(), so it can not be in the if_proj list.
2364 continue;
2365 }
2366 }
2368 Node* test = iff->in(1);
2369 if (!test->is_Bool()){ //Conv2B, ...
2370 continue;
2371 }
2372 BoolNode* bol = test->as_Bool();
2373 if (invar.is_invariant(bol)) {
2374 // Invariant test
2375 new_predicate_proj = create_new_if_for_predicate(predicate_proj, NULL,
2376 Deoptimization::Reason_predicate);
2377 Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0);
2378 BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool();
2380 // Negate test if necessary
2381 bool negated = false;
2382 if (proj->_con != predicate_proj->_con) {
2383 new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate());
2384 register_new_node(new_predicate_bol, ctrl);
2385 negated = true;
2386 }
2387 IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If();
2388 _igvn.hash_delete(new_predicate_iff);
2389 new_predicate_iff->set_req(1, new_predicate_bol);
2390 #ifndef PRODUCT
2391 if (TraceLoopPredicate) {
2392 tty->print("Predicate invariant if%s: %d ", negated ? " negated" : "", new_predicate_iff->_idx);
2393 loop->dump_head();
2394 } else if (TraceLoopOpts) {
2395 tty->print("Predicate IC ");
2396 loop->dump_head();
2397 }
2398 #endif
2399 } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) {
2400 assert(proj->_con == predicate_proj->_con, "must match");
2402 // Range check for counted loops
2403 const Node* cmp = bol->in(1)->as_Cmp();
2404 Node* idx = cmp->in(1);
2405 assert(!invar.is_invariant(idx), "index is variant");
2406 assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be");
2407 Node* rng = cmp->in(2);
2408 assert(invar.is_invariant(rng), "range must be invariant");
2409 int scale = 1;
2410 Node* offset = zero;
2411 bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset);
2412 assert(ok, "must be index expression");
2414 Node* init = cl->init_trip();
2415 Node* limit = cl->limit();
2416 Node* stride = cl->stride();
2418 // Build if's for the upper and lower bound tests. The
2419 // lower_bound test will dominate the upper bound test and all
2420 // cloned or created nodes will use the lower bound test as
2421 // their declared control.
2422 ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
2423 ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj, NULL, Deoptimization::Reason_predicate);
2424 assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate");
2425 Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0);
2427 // Perform cloning to keep Invariance state correct since the
2428 // late schedule will place invariant things in the loop.
2429 rng = invar.clone(rng, ctrl);
2430 if (offset && offset != zero) {
2431 assert(invar.is_invariant(offset), "offset must be loop invariant");
2432 offset = invar.clone(offset, ctrl);
2433 }
2435 // Test the lower bound
2436 Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false);
2437 IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If();
2438 _igvn.hash_delete(lower_bound_iff);
2439 lower_bound_iff->set_req(1, lower_bound_bol);
2440 if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx);
2442 // Test the upper bound
2443 Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true);
2444 IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If();
2445 _igvn.hash_delete(upper_bound_iff);
2446 upper_bound_iff->set_req(1, upper_bound_bol);
2447 if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx);
2449 // Fall through into rest of the clean up code which will move
2450 // any dependent nodes onto the upper bound test.
2451 new_predicate_proj = upper_bound_proj;
2453 #ifndef PRODUCT
2454 if (TraceLoopOpts && !TraceLoopPredicate) {
2455 tty->print("Predicate RC ");
2456 loop->dump_head();
2457 }
2458 #endif
2459 } else {
2460 // Loop variant check (for example, range check in non-counted loop)
2461 // with uncommon trap.
2462 continue;
2463 }
2464 assert(new_predicate_proj != NULL, "sanity");
2465 // Success - attach condition (new_predicate_bol) to predicate if
2466 invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate
2468 // Eliminate the old If in the loop body
2469 dominated_by( new_predicate_proj, iff, proj->_con != new_predicate_proj->_con );
2471 hoisted = true;
2472 C->set_major_progress();
2473 } // end while
2475 #ifndef PRODUCT
2476 // report that the loop predication has been actually performed
2477 // for this loop
2478 if (TraceLoopPredicate && hoisted) {
2479 tty->print("Loop Predication Performed:");
2480 loop->dump_head();
2481 }
2482 #endif
2484 return hoisted;
2485 }
2487 //------------------------------loop_predication--------------------------------
2488 // driver routine for loop predication optimization
2489 bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) {
2490 bool hoisted = false;
2491 // Recursively promote predicates
2492 if ( _child ) {
2493 hoisted = _child->loop_predication( phase);
2494 }
2496 // self
2497 if (!_irreducible && !tail()->is_top()) {
2498 hoisted |= phase->loop_predication_impl(this);
2499 }
2501 if ( _next ) { //sibling
2502 hoisted |= _next->loop_predication( phase);
2503 }
2505 return hoisted;
2506 }
2509 // Process all the loops in the loop tree and replace any fill
2510 // patterns with an intrisc version.
2511 bool PhaseIdealLoop::do_intrinsify_fill() {
2512 bool changed = false;
2513 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
2514 IdealLoopTree* lpt = iter.current();
2515 changed |= intrinsify_fill(lpt);
2516 }
2517 return changed;
2518 }
2521 // Examine an inner loop looking for a a single store of an invariant
2522 // value in a unit stride loop,
2523 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
2524 Node*& shift, Node*& con) {
2525 const char* msg = NULL;
2526 Node* msg_node = NULL;
2528 store_value = NULL;
2529 con = NULL;
2530 shift = NULL;
2532 // Process the loop looking for stores. If there are multiple
2533 // stores or extra control flow give at this point.
2534 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2535 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2536 Node* n = lpt->_body.at(i);
2537 if (n->outcnt() == 0) continue; // Ignore dead
2538 if (n->is_Store()) {
2539 if (store != NULL) {
2540 msg = "multiple stores";
2541 break;
2542 }
2543 int opc = n->Opcode();
2544 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) {
2545 msg = "oop fills not handled";
2546 break;
2547 }
2548 Node* value = n->in(MemNode::ValueIn);
2549 if (!lpt->is_invariant(value)) {
2550 msg = "variant store value";
2551 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
2552 msg = "not array address";
2553 }
2554 store = n;
2555 store_value = value;
2556 } else if (n->is_If() && n != head->loopexit()) {
2557 msg = "extra control flow";
2558 msg_node = n;
2559 }
2560 }
2562 if (store == NULL) {
2563 // No store in loop
2564 return false;
2565 }
2567 if (msg == NULL && head->stride_con() != 1) {
2568 // could handle negative strides too
2569 if (head->stride_con() < 0) {
2570 msg = "negative stride";
2571 } else {
2572 msg = "non-unit stride";
2573 }
2574 }
2576 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
2577 msg = "can't handle store address";
2578 msg_node = store->in(MemNode::Address);
2579 }
2581 if (msg == NULL &&
2582 (!store->in(MemNode::Memory)->is_Phi() ||
2583 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
2584 msg = "store memory isn't proper phi";
2585 msg_node = store->in(MemNode::Memory);
2586 }
2588 // Make sure there is an appropriate fill routine
2589 BasicType t = store->as_Mem()->memory_type();
2590 const char* fill_name;
2591 if (msg == NULL &&
2592 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
2593 msg = "unsupported store";
2594 msg_node = store;
2595 }
2597 if (msg != NULL) {
2598 #ifndef PRODUCT
2599 if (TraceOptimizeFill) {
2600 tty->print_cr("not fill intrinsic candidate: %s", msg);
2601 if (msg_node != NULL) msg_node->dump();
2602 }
2603 #endif
2604 return false;
2605 }
2607 // Make sure the address expression can be handled. It should be
2608 // head->phi * elsize + con. head->phi might have a ConvI2L.
2609 Node* elements[4];
2610 Node* conv = NULL;
2611 bool found_index = false;
2612 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
2613 for (int e = 0; e < count; e++) {
2614 Node* n = elements[e];
2615 if (n->is_Con() && con == NULL) {
2616 con = n;
2617 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
2618 Node* value = n->in(1);
2619 #ifdef _LP64
2620 if (value->Opcode() == Op_ConvI2L) {
2621 conv = value;
2622 value = value->in(1);
2623 }
2624 #endif
2625 if (value != head->phi()) {
2626 msg = "unhandled shift in address";
2627 } else {
2628 found_index = true;
2629 shift = n;
2630 assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match");
2631 }
2632 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
2633 if (n->in(1) == head->phi()) {
2634 found_index = true;
2635 conv = n;
2636 } else {
2637 msg = "unhandled input to ConvI2L";
2638 }
2639 } else if (n == head->phi()) {
2640 // no shift, check below for allowed cases
2641 found_index = true;
2642 } else {
2643 msg = "unhandled node in address";
2644 msg_node = n;
2645 }
2646 }
2648 if (count == -1) {
2649 msg = "malformed address expression";
2650 msg_node = store;
2651 }
2653 if (!found_index) {
2654 msg = "missing use of index";
2655 }
2657 // byte sized items won't have a shift
2658 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
2659 msg = "can't find shift";
2660 msg_node = store;
2661 }
2663 if (msg != NULL) {
2664 #ifndef PRODUCT
2665 if (TraceOptimizeFill) {
2666 tty->print_cr("not fill intrinsic: %s", msg);
2667 if (msg_node != NULL) msg_node->dump();
2668 }
2669 #endif
2670 return false;
2671 }
2673 // No make sure all the other nodes in the loop can be handled
2674 VectorSet ok(Thread::current()->resource_area());
2676 // store related values are ok
2677 ok.set(store->_idx);
2678 ok.set(store->in(MemNode::Memory)->_idx);
2680 // Loop structure is ok
2681 ok.set(head->_idx);
2682 ok.set(head->loopexit()->_idx);
2683 ok.set(head->phi()->_idx);
2684 ok.set(head->incr()->_idx);
2685 ok.set(head->loopexit()->cmp_node()->_idx);
2686 ok.set(head->loopexit()->in(1)->_idx);
2688 // Address elements are ok
2689 if (con) ok.set(con->_idx);
2690 if (shift) ok.set(shift->_idx);
2691 if (conv) ok.set(conv->_idx);
2693 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2694 Node* n = lpt->_body.at(i);
2695 if (n->outcnt() == 0) continue; // Ignore dead
2696 if (ok.test(n->_idx)) continue;
2697 // Backedge projection is ok
2698 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
2699 if (!n->is_AddP()) {
2700 msg = "unhandled node";
2701 msg_node = n;
2702 break;
2703 }
2704 }
2706 // Make sure no unexpected values are used outside the loop
2707 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
2708 Node* n = lpt->_body.at(i);
2709 // These values can be replaced with other nodes if they are used
2710 // outside the loop.
2711 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
2712 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
2713 Node* use = iter.get();
2714 if (!lpt->_body.contains(use)) {
2715 msg = "node is used outside loop";
2716 // lpt->_body.dump();
2717 msg_node = n;
2718 break;
2719 }
2720 }
2721 }
2723 #ifdef ASSERT
2724 if (TraceOptimizeFill) {
2725 if (msg != NULL) {
2726 tty->print_cr("no fill intrinsic: %s", msg);
2727 if (msg_node != NULL) msg_node->dump();
2728 } else {
2729 tty->print_cr("fill intrinsic for:");
2730 }
2731 store->dump();
2732 if (Verbose) {
2733 lpt->_body.dump();
2734 }
2735 }
2736 #endif
2738 return msg == NULL;
2739 }
2743 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
2744 // Only for counted inner loops
2745 if (!lpt->is_counted() || !lpt->is_inner()) {
2746 return false;
2747 }
2749 // Must have constant stride
2750 CountedLoopNode* head = lpt->_head->as_CountedLoop();
2751 if (!head->stride_is_con() || !head->is_normal_loop()) {
2752 return false;
2753 }
2755 // Check that the body only contains a store of a loop invariant
2756 // value that is indexed by the loop phi.
2757 Node* store = NULL;
2758 Node* store_value = NULL;
2759 Node* shift = NULL;
2760 Node* offset = NULL;
2761 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
2762 return false;
2763 }
2765 // Now replace the whole loop body by a call to a fill routine that
2766 // covers the same region as the loop.
2767 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
2769 // Build an expression for the beginning of the copy region
2770 Node* index = head->init_trip();
2771 #ifdef _LP64
2772 index = new (C, 2) ConvI2LNode(index);
2773 _igvn.register_new_node_with_optimizer(index);
2774 #endif
2775 if (shift != NULL) {
2776 // byte arrays don't require a shift but others do.
2777 index = new (C, 3) LShiftXNode(index, shift->in(2));
2778 _igvn.register_new_node_with_optimizer(index);
2779 }
2780 index = new (C, 4) AddPNode(base, base, index);
2781 _igvn.register_new_node_with_optimizer(index);
2782 Node* from = new (C, 4) AddPNode(base, index, offset);
2783 _igvn.register_new_node_with_optimizer(from);
2784 // Compute the number of elements to copy
2785 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip());
2786 _igvn.register_new_node_with_optimizer(len);
2788 BasicType t = store->as_Mem()->memory_type();
2789 bool aligned = false;
2790 if (offset != NULL && head->init_trip()->is_Con()) {
2791 int element_size = type2aelembytes(t);
2792 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
2793 }
2795 // Build a call to the fill routine
2796 const char* fill_name;
2797 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
2798 assert(fill != NULL, "what?");
2800 // Convert float/double to int/long for fill routines
2801 if (t == T_FLOAT) {
2802 store_value = new (C, 2) MoveF2INode(store_value);
2803 _igvn.register_new_node_with_optimizer(store_value);
2804 } else if (t == T_DOUBLE) {
2805 store_value = new (C, 2) MoveD2LNode(store_value);
2806 _igvn.register_new_node_with_optimizer(store_value);
2807 }
2809 Node* mem_phi = store->in(MemNode::Memory);
2810 Node* result_ctrl;
2811 Node* result_mem;
2812 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
2813 int size = call_type->domain()->cnt();
2814 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill,
2815 fill_name, TypeAryPtr::get_array_body_type(t));
2816 call->init_req(TypeFunc::Parms+0, from);
2817 call->init_req(TypeFunc::Parms+1, store_value);
2818 #ifdef _LP64
2819 len = new (C, 2) ConvI2LNode(len);
2820 _igvn.register_new_node_with_optimizer(len);
2821 #endif
2822 call->init_req(TypeFunc::Parms+2, len);
2823 #ifdef _LP64
2824 call->init_req(TypeFunc::Parms+3, C->top());
2825 #endif
2826 call->init_req( TypeFunc::Control, head->init_control());
2827 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
2828 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
2829 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
2830 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
2831 _igvn.register_new_node_with_optimizer(call);
2832 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
2833 _igvn.register_new_node_with_optimizer(result_ctrl);
2834 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory);
2835 _igvn.register_new_node_with_optimizer(result_mem);
2837 // If this fill is tightly coupled to an allocation and overwrites
2838 // the whole body, allow it to take over the zeroing.
2839 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
2840 if (alloc != NULL && alloc->is_AllocateArray()) {
2841 Node* length = alloc->as_AllocateArray()->Ideal_length();
2842 if (head->limit() == length &&
2843 head->init_trip() == _igvn.intcon(0)) {
2844 if (TraceOptimizeFill) {
2845 tty->print_cr("Eliminated zeroing in allocation");
2846 }
2847 alloc->maybe_set_complete(&_igvn);
2848 } else {
2849 #ifdef ASSERT
2850 if (TraceOptimizeFill) {
2851 tty->print_cr("filling array but bounds don't match");
2852 alloc->dump();
2853 head->init_trip()->dump();
2854 head->limit()->dump();
2855 length->dump();
2856 }
2857 #endif
2858 }
2859 }
2861 // Redirect the old control and memory edges that are outside the loop.
2862 Node* exit = head->loopexit()->proj_out(0);
2863 // Sometimes the memory phi of the head is used as the outgoing
2864 // state of the loop. It's safe in this case to replace it with the
2865 // result_mem.
2866 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
2867 _igvn.replace_node(exit, result_ctrl);
2868 _igvn.replace_node(store, result_mem);
2869 // Any uses the increment outside of the loop become the loop limit.
2870 _igvn.replace_node(head->incr(), head->limit());
2872 // Disconnect the head from the loop.
2873 for (uint i = 0; i < lpt->_body.size(); i++) {
2874 Node* n = lpt->_body.at(i);
2875 _igvn.replace_node(n, C->top());
2876 }
2878 return true;
2879 }